Infrared remote sensing image super-resolution network by integration of dense connection and multi-attention mechanism

Fig. 2. Attention feature extract block（AFEB）

Fig. 3. Structure of Swin Transformer block（STB）

Fig. 4. Structure of multi-scale channel attention（MSCA）

Fig. 5. Comparison of reconstructed images by different super resolution algorithms on the ISDD test set: (a) HR (PSNR/SSIM); (b) Bicubic (36.10/0.9572); (c) ESRGAN (34.32/0.9340); (d) CRAFT-SR (39.54/0.9766); (e) SAMFN (39.51/0.9758);（f）DSSR（38.27/0.9695）；（g）TransENet（38.48/0.9648）；（h）ours（39.55/0.9769）

Fig. 6. Comparison of reconstructed images by different super resolution algorithms on the NUDT-SIRST-Sea test set:：（a）HR（PSNR/SSIM）；（b）Bicubic（30.04/0.8152）；（c）ESRGAN（29.46/0.8164）；（d）CRAFT-SR（31.69/0.8789）；（e）SAMFN（31.63/0.8736）；（f）DSSR（31.40/0.8654）；（g）TransENet（31.15/0.8564）；（h）ours（31.77/0.8818）

Table 1. Ablation results of ISDD

Table 1. Ablation results of ISDD

	Model1	Model2	Model3	Model4	Model5
Dense block		√	√	√	√
MSCA			√	√	√
Joint loss				√	√
Discriminator					√
PSNR/dB	46.5374	46.8726	46.9220	46.9741	47.0499
SSIM	0.9764	0.9785	0.9789	0.9795	0.9801
FID	12.4895	11.1970	10.9424	9.9607	9.7600

Table 2. Ablation results of NUDT-SIRST-Sea

Table 2. Ablation results of NUDT-SIRST-Sea

	Model1	Model2	Model3	Model4	Model5
Dense block		√	√	√	√
MSCA			√	√	√
Joint loss				√	√
Discriminator					√
PSNR/dB	41.7496	41.7587	41.7732	41.7789	41.7804
SSIM	0.9415	0.9419	0.9427	0.9447	0.9450
FID	16.4868	15.8964	15.7469	15.1687	14.5104

Table 3. Impact of each submodule on the computational complexity of the model
View table
View in Article
Table 3. Impact of each submodule on the computational complexity of the model
Model1 Model2 Model3 Model4 Model5
可训练参数量/10⁶ 3.2576 3.2576 3.2598 3.2598 7.0915
FLOPs/10⁹ 58.3334 58.3334 58.3512 58.3512 59.1132
平均推理时间/s 0.0344 0.0346 0.0358 0.0362 0.0362

Table 4. Comparison results of different super resolution algorithms on the ISDD test set

Table 4. Comparison results of different super resolution algorithms on the ISDD test set

评估指标	HR	bicubic	ESRGAN	CRAFT-SR	SAFMN	DSSR	TransENet	Ours
PSNR	\	46.2894	45.3329	46.9941	46.9751	46.3847	44.8646	47.0499
SSIM	\	0.9751	0.9692	0.9796	0.9784	0.9771	0.9735	0.9801
FID	\	18.9760	16.3093	9.9029	9.9545	11.3072	19.5806	9.7600
mAP50	0.950	0.940	0.925	0.938	0.944	0.934	0.943	0.947
mAP50-95	0.536	0.525	0.507	0.523	0.527	0.515	0.525	0.530

Table 5. Comparison results of different super resolution algorithms on the NUDT-SIRST-Sea test set

Table 5. Comparison results of different super resolution algorithms on the NUDT-SIRST-Sea test set

评估指标	HR	bicubic	ESRGAN	CRAFT-SR	SAFMN	DSSR	TransENet	Ours
PSNR	\	41.1219	39.1022	41.7775	41.6659	41.6141	40.7012	41.7804
SSIM	\	0.9301	0.9243	0.9435	0.9430	0.9413	0.9373	0.9450
FID	\	33.0765	41.1734	14.7453	14.6553	15.0255	19.3459	14.5104
mAP50	0.667	0.528	0.552	0.555	0.551	0.552	0.554	0.578
mAP50-95	0.425	0.298	0.320	0.332	0.330	0.324	0.323	0.340

Table 6. Comparison of computational complexity of different super-resolution algorithms